Abnormal Dynamic Detection of Space Targets from ISAR Sequences with LSTM Regression

In both military and civilian contexts, accurate detection of radar attitude deviations using ISAR is crucial but challenging due to the incomplete target edges and missing electromagnetic scattering components. Available dynamic detection of space targets transforms detection into binary classification, which faces an imbalanced dataset and an incomplete repre-sentation of all anomalous samples. In this paper, we propose an abnormal dynamic detection method with small amounts of anomalies. An LSTM architecture is first constructed for the geometrical feature flow regression for stabilized space targets. Then, a small number of anomalies and normal samples are used to calculate the value of the mean absolute error (MAE) threshold to distinguish anomalies from normal samples using a statistical histogram. In this way, abnormal dynamic detection can be accomplished by comparing MAE to the threshold. Experiments validate the effectiveness of the proposal with limited anomalies.